Meta has announced the launch of Llama 4, a new family of AI models that boast enhanced multimodal capabilities, allowing them to process and understand large amounts of unlabeled text, image, and video data. The release comes as a response to the success of open models from Chinese AI lab DeepSeek, which reportedly prompted Meta to accelerate the development of its Llama models.

The Llama 4 collection consists of four new models: Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. While Scout and Maverick are openly available on Llama.com and through Meta's partners, including Hugging Face, Behemoth is still in training. Notably, Meta AI, the company's AI-powered assistant across apps like WhatsApp, Messenger, and Instagram, has been updated to use Llama 4 in 40 countries, with multimodal features limited to the U.S. in English for now.

However, some developers may take issue with the Llama 4 license, which prohibits users and companies "domiciled" or with a "principal place of business" in the EU from using or distributing the models. This restriction is likely due to governance requirements imposed by the region's AI and data privacy laws. Additionally, companies with more than 700 million monthly active users must request a special license from Meta, which can be granted or denied at the company's discretion.

Meta claims that Llama 4 marks a new era for the Llama ecosystem, with the models utilizing a mixture of experts (MoE) architecture, which is more computationally efficient for training and answering queries. Maverick, for example, has 400 billion total parameters, but only 17 billion active parameters across 128 "experts." Scout, on the other hand, has 17 billion active parameters, 16 experts, and 109 billion total parameters.

According to Meta's internal testing, Maverick exceeds models like OpenAI's GPT-4o and Google's Gemini 2.0 on certain coding, reasoning, multilingual, long-context, and image benchmarks. However, it doesn't quite measure up to more capable recent models like Google's Gemini 2.5 Pro, Anthropic's Claude 3.7 Sonnet, and OpenAI's GPT-4.5. Scout's strengths lie in tasks like document summarization and reasoning over large codebases, with a unique feature being its very large context window of 10 million tokens.

Meta's unreleased Behemoth model will require even more powerful hardware, with 288 billion active parameters, 16 experts, and nearly two trillion total parameters. Internal benchmarking has Behemoth outperforming GPT-4.5, Claude 3.7 Sonnet, and Gemini 2.0 Pro on several evaluations measuring STEM skills like math problem solving.

Interestingly, Meta has tuned all of its Llama 4 models to refuse to answer "contentious" questions less often, with the goal of providing more balanced and factual responses. This move comes as some accuse AI models of political bias, with White House allies claiming that AI chatbots censor conservative viewpoints. However, experts agree that bias in AI is an intractable technical problem, and companies like OpenAI have also adjusted their models to answer more questions on controversial political subjects.

In conclusion, the release of Llama 4 marks a significant milestone in the development of AI models, with Meta's new collection boasting enhanced multimodal capabilities and improved performance. However, the licensing restrictions and limitations may pose challenges for some developers and users. As the AI landscape continues to evolve, it remains to be seen how these models will be adopted and utilized in various applications.